# opencode-config > A personal OpenCode configuration that turns Claude into a self-learning multi-agent swarm system with parallel workers, outcome-based pattern learning, and cross-agent memory. opencode-config is a personal OpenCode configuration by Joel Hooks that transforms the OpenCode AI coding environment into a multi-agent swarm system. It is built on top of `joelhooks/swarmtools` and runs entirely inside OpenCode — it is not a standalone tool. The repository has accumulated over 500 GitHub stars since its creation in December 2025, signaling strong community interest in the approach. ## What It Is opencode-config is a configuration layer for [OpenCode](https://opencode.ai) that wires together a coordinator-worker agent architecture, outcome-based learning, cross-agent session memory, and a suite of custom MCP tools. When you describe a task inside OpenCode, the system decomposes it into subtasks, spawns parallel worker agents, tracks which strategies succeed or fail, and adapts over time. Anti-patterns get detected and inverted; proven patterns get promoted. Confidence in patterns decays on a 90-day half-life unless revalidated. ## How the Swarm Learns The core innovation is an outcome-based learning pipeline that feeds every task execution back into a pattern maturity system: - **Fast + success** → pattern gets promoted toward "proven" - **Slow + retries + errors** → pattern gets flagged - **>60% failure rate** → auto-inverted to anti-pattern - **90-day half-life** → confidence decays unless revalidated Patterns move through four maturity stages: `candidate → established → proven → deprecated`. This means the swarm improves with use rather than repeating the same mistakes. ## Architecture: Coordinator vs. Worker The system separates expensive, long-lived coordinator context from disposable worker context. The coordinator (running on Claude Sonnet) never edits code directly — it decomposes tasks, orchestrates workers, and monitors progress. Workers get focused, isolated context windows, execute their assigned tasks, checkpoint progress, and report learning signals back. The README claims this architecture produces a 70% cost reduction compared to a single-agent approach, while also improving recovery and generating better learning signals. ## Cross-Agent Memory with CASS and Hivemind Two memory systems extend the swarm's knowledge across sessions and agents: - **CASS (Cross-Agent Session Search)** indexes conversation histories from Claude Code, Codex, Cursor, Gemini, Aider, ChatGPT, Cline, OpenCode, Amp, and Pi-Agent, enabling semantic and full-text search across all past AI sessions before solving a new problem. - **Hivemind (Semantic Memory)** persists architectural decisions, debugging breakthroughs, and project-specific gotchas using vector search backed by the `nomic-embed-text` embedding model via Ollama. ## Custom Tools and MCP Servers The config ships 12 custom MCP tools and integrates several MCP servers: - **UBS (Ultimate Bug Scanner)** — multi-language static analysis covering JS/TS, Python, C++, Rust, Go, Java, and Ruby, checking for null safety, XSS, injection, async race conditions, and memory leaks - **Hive** — git-backed work tracker for atomic epic and subtask creation with status tracking - **Agent Mail** — file reservation and message-passing system to prevent edit conflicts between parallel agents - **repo-autopsy** — deep GitHub repo analysis using AST grep, blame, hotspots, and secret detection - **pdf-brain** — PDF and Markdown knowledge base with URL support MCP servers include next-devtools, chrome-devtools, context7 (library docs), fetch, Snyk (security scanning), and Kernel (cloud browser automation via Playwright). ## Setup Path Installation requires OpenCode itself (via install script or Homebrew), then cloning this repo into `~/.config/opencode` and running `bun install`. Two npm CLIs are required as agent backends: `opencode-swarm-plugin` (swarm orchestration) and optionally `cass-search` (cross-agent session search). Ollama with the `nomic-embed-text` model is required for semantic memory features. Once installed, running `/init` inside OpenCode on a target repo generates an `AGENTS.md` workflow file. Swarms are triggered with `/swarm "describe your task"` from within OpenCode. ## Update: Active Development as of Early 2026 The repository was created in December 2025 and last pushed in January 2026, with the GitHub metadata showing a last-updated timestamp of August 2026. The swarm CLI is pinned at version 0.30.0, cass at 0.1.35, and Ollama at 0.13.1. The project is MIT-licensed and actively maintained by Joel Hooks as a personal configuration that others can clone and adapt. ## Features - Multi-agent swarm orchestration with parallel workers - Outcome-based pattern learning with confidence decay - Cross-agent session search (CASS) across 10+ AI tools - Semantic memory (Hivemind) with vector search via Ollama - UBS multi-language bug scanner (JS/TS, Python, Rust, Go, etc.) - Coordinator-worker architecture for cost optimization - Git-backed work tracker (Hive) with epic/subtask management - Agent Mail file reservation to prevent edit conflicts - 25 slash commands (/swarm, /debug, /iterate, etc.) - 7 injectable skill packages (testing-patterns, cli-builder, etc.) - MCP server integrations (next-devtools, chrome-devtools, Snyk, Kernel) - Knowledge files for TDD, TypeScript, Next.js, Effect-TS patterns - Anti-pattern detection and auto-inversion at >60% failure rate - Cross-session learning with 90-day confidence half-life - AGENTS.md workflow file generation via /init command ## Integrations OpenCode, Claude (Sonnet, Haiku), Ollama (nomic-embed-text), Claude Code, Cursor, Codex, Gemini, Aider, ChatGPT, Cline, Amp, Next.js, Snyk, Playwright (via Kernel), context7, GitHub API ## Platforms MACOS, API, CLI ## Pricing Open Source ## Version swarm 0.30.0 / cass 0.1.35 ## Links - Website: https://github.com/joelhooks/opencode-config - Documentation: https://github.com/joelhooks/opencode-config#readme - Repository: https://github.com/joelhooks/opencode-config - EveryDev.ai: https://www.everydev.ai/tools/opencode-config